This is why you can’t build your own RTX Spark PC
· pcworld.com · popularity 21/100 · 0 similar stories today · Home, Garden & DIY (Home, Garden & DIY)
Read the original at pcworld.com →
Summary
The Asus ProArt P16 RTX Spark has 128GB of memory. So does my desktop, basically. My desktop also has a GeForce RTX 5090 , the fastest consumer graphics card Nvidia has ever made, and one that’s an absolute monster for AI tasks. So, why did the laptop finish my massive local AI workload almost 20 minutes faster? Because my desktop has all that memory in the wrong damned places. I’ve been putting the staggeringly expensive $7,000 Asus ProArt P16 laptop through some unusual tests for my AI-focused RTX Spark review , including running a sprawling real-world research workload through OpenAI’s gigantic 120-billion-parameter gpt-oss model. Rather than asking an AI to write a poem about Jensen Huang or make me a picture of a cat on a motorcycle, I wanted to see whether this specialized hardware could handle the sort of stupidly ambitious local AI work I’d normally reserve for cloud models. It can. More interestingly, my monster desktop can too. Just much, much slower. For one frozen comparison, both systems chewed through the exact same workload, generating 472,084 tokens along the way. The ProArt P16 RTX Spark finished in 33 minutes and 13 seconds. My desktop needed 52 minutes and 51 seconds. That’s a 19-minute, 38-second ass-kicking delivered to a fully kitted, fire-breathing high-end desktop packing an RTX 5090. The reason gets to the heart of what makes these new high-memory AI systems so interesting — and why you can’t simply build your own equivalent with conventional enthusiast PC parts. 128GB isn’t always 128GB The ProArt P16 RTX Spark packs 128GB of unified LPDDR5X memory, with roughly 96GB available to the GPU during my testing. My desktop pairs its RTX 5090’s screaming-fast 32GB of GDDR7 memory with 96GB of conventional system RAM. Look at those numbers again. The two machines have virtually the same total amount of memory. They’ve just flipped where most of it is available. Nvidia’s ferocious GeForce RX 5090 in a standard desktop system (not mine). Adam Patrick Murray / Foundry For traditional PC workloads, my desktop’s arrangement makes perfect sense. The CPU gets a huge pool of system memory. The graphics card gets a smaller pool of ultra-fast dedicated memory. Games love it. Most GPU workloads love it. And the RTX 5090 is an absolute monster when the work fits inside that 32GB frame buffer. A 120-billion-parameter AI model doesn’t. My desktop can still run the model by offloading what won’t fit in the RTX 5090’s 32GB of VRAM to system memory. But that’s very different from keeping far more of the workload directly accessible to the GPU in a large unified memory pool. The RTX Spark machine was designed around that sort of high-capacity local AI workload. Its unified memory architecture gives the GPU access to vastly more memory than a conventional consumer graphics card. For a model this large, that can matter enormously. To be clear, this isn’t a controlled memory benchmark. These are wildly different systems, and GPU architecture, memory bandwidth, software behavior, and other differences can all affect performance. But my RTX 5090 runs headfirst into a very obvious constraint here: It has 32GB of dedicated VRAM, while the Spark system can make a much larger chunk of its 128GB unified pool available to the GPU. So why can’t I just build one? This is not remotely the same as memory directly allocated to a GPU despite costing a fortune in 2026. G.Skill There’s no obvious DIY fix. I can’t walk into Micro Center and buy a 96GB RTX 5090. Adding another 128GB of DDR5 to my motherboard wouldn’t turn it into GPU memory. Throwing an even faster desktop processor at the problem doesn’t change the 5090’s 32GB VRAM capacity. You can attack these problems with vastly more exotic hardware, multiple accelerators, or other compromises, of course. But at that point you aren’t building the obvious enthusiast-PC equivalent of an RTX Spark system anymore. Local AI changes the PC spec sheet That’s what opened my eyes the most during this intensive two weeks of real-world testing. I’ve spent decades looking at PC spec sheets, and memory capacity has always seemed like one of the easiest numbers to understand. Eight gigs. Sixteen gigs. Thirty-two gigs. More is more. Local AI makes that deceptively complicated. My desktop and the ProArt P16 RTX Spark both have roughly 128GB of memory available. One has the vastly faster discrete GPU. Yet the other crushed it in this particular test. The comparison isn’t proof that unified memory alone caused the difference, but it exposes something important: Total memory capacity no longer tells you nearly enough. Where that memory is available, how quickly it can be accessed, and what hardware can use it all matter. I started this experiment wondering whether I could essentially build my own RTX Spark with off-the-shelf PC hardware. It turns out I already sort of had in my unending quest for MOAR PC GAMING FRAMES. I had the memory, the ludicrously powerful GPU, even roughly the same total capacity! I just had all the memory in the wrong damned places.
Found in: PCWorld. Rating basis: single outlet; outlet tier 2 (unlisted); 1 feeds; feed position 2.